Most neural network libraries, including PyTorch, scikit and Keras, have built-in MNIST datasets. However, working with pre-built MNIST datasets has two big problems.